Date of Award

Summer 2021

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

Committee Director

Nikos Chrisochoides

Committee Member

Michael A. Park

Committee Member

Andrey Chernikov

Committee Member

Desh Ranjan

Committee Member

Charles E. Hyde

Abstract

Finite-element methods are a critical component of the design and analysis procedures of many (bio-)engineering applications. Mesh adaptation is one of the most crucial components since it discretizes the physics of the application at a relatively low cost to the solver. Highly scalable parallel mesh adaptation methods for High-Performance Computing (HPC) are essential to meet the ever-growing demand for higher fidelity simulations. Moreover, the continuous growth of the complexity of the HPC systems requires a systematic approach to exploit their full potential. Anisotropic mesh adaptation captures features of the solution at multiple scales while, minimizing the required number of elements. However, it also introduces new challenges on top of mesh generation. Also, the increased complexity of the targeted cases requires departing from traditional surface-constrained approaches to utilizing CAD (Computer-Aided Design) kernels. Alongside the functionality requirements, is the need of taking advantage of the ubiquitous multi-core machines. More importantly, the parallel implementation needs to handle the ever-increasing complexity of the mesh adaptation code.

In this work, we develop a parallel mesh adaptation method that utilizes a metric-based approach for generating anisotropic meshes. Moreover, we enhance our method by interfacing with a CAD kernel, thus enabling its use on complex geometries. We evaluate our method both with fixed-resolution benchmarks and within a simulation pipeline, where the resolution of the discretization increases incrementally. With the Telescopic Approach for scalable mesh generation as a guide, we propose a parallel method at the node (multi-core) for mesh adaptation that is expected to scale up efficiently to the upcoming exascale machines. To facilitate an effective implementation, we introduce an abstract layer between the application and the runtime system that enables the use of task-based parallelism for concurrent mesh operations. Our evaluation indicates results comparable to state-of-the-art methods for fixed-resolution meshes both in terms of performance and quality. The integration with an adaptive pipeline offers promising results for the capability of the proposed method to function as part of an adaptive simulation. Moreover, our abstract tasking layer allows the separation of different aspects of the implementation without any impact on the functionality of the method.

DOI

10.25777/ask4-r595

ORCID

0000-0002-0656-9631

Share

COinS